Claude for finance teams

Most bad AI finance numbers aren't invented from nothing. They're a real figure answering a different question: cash basis when the pack is accrual, a payout counted in the wrong month, MRR quoted as revenue. An assistant that doesn't know your conventions picks one and doesn't say so. Write the conventions down once, in the project, and Claude can say which one each figure uses.

The definitions themselves are on monthly recurring revenue, days sales outstanding and budget variance analysis. The project instructions in the setup below turn this table into rules Claude follows.

Most finance teams don't need a new platform. Claude sits in the middle: a project holds the context above, a skill holds your definitions, and each vendor's own connector links it to the ledger, billing and CRM that already hold the numbers. Count sits beside it and checks the answers. On NetSuite, Oracle's AI Connector Service connects Claude through a non-administrator role; Sage has no official connector yet, so its reports go in as CSV.

What Anthropic offers finance teams

Open an inherited three-statement model, click the Claude add-in, and type "Walk me through how the revenue number in cell C42 is calculated." Claude answers with cell-level citations you can click to jump to each input (Anthropic, Use Claude for Excel). That add-in is the centre of what Anthropic offers finance teams in October 2026.

Prices from claude.com/pricing; the rest from Anthropic's docs and launch posts, checked 22 October 2026.

Most of the marketing is aimed at banks and asset managers: pitch books, credit memos, KYC screening. For a corporate finance or FP&A team the useful parts are narrower. The Excel add-in, a project holding your definitions, and read-only connectors to your own ledger, billing and warehouse cover most of the work, and a Team plan gets you all three.

If your company runs ChatGPT instead.

Xero, QuickBooks and Stripe each offer a ChatGPT plugin too (Xero Blog; Intuit; Stripe Docs), and ChatGPT Business costs the same as a Claude Team seat ($25 a user monthly, $20 annually). The project instructions below work as a ChatGPT project's instructions. For workbooks, OpenAI has its own ChatGPT for Excel add-in. The full ChatGPT setup is in ChatGPT for finance.

You don't need to be technical for most of this. The steps marked You happen in Claude's settings and take about half an hour between them. Xero, QuickBooks and Stripe all sign in from Claude, so IT only gets involved if you run NetSuite or want Claude for Excel deployed company-wide. Step 1 gives the project the context from the top of this page, so every workflow below starts from your conventions instead of a guess.

Technical steps for NetSuite (for your NetSuite administrator)

• Install the MCP Standard Tools SuiteApp. Create a non-administrator role with the MCP Server Connection and OAuth 2.0 Access Tokens permissions, plus REST Web Services for tools that need it (NetSuite AI Connector Service).
• Give the role view access only to the records the finance team reads. The connector can create and update records, so the role is where reading-only is enforced.
• In Claude, add a custom connector with the URL https://<accountid>.suitetalk.api.netsuite.com/services/mcp/v1/all. Oracle notes that without /all at the end the connection appears disconnected.

Keep Claude reading, not changing, until you trust it.

Xero's connector can only read. QuickBooks' can create, send and delete invoices, Stripe's can write through the API (Stripe asks a person to confirm refunds), and Chargebee, HubSpot and Claude for Excel can change data too. When Claude asks to change something, approve it one at a time and never choose "Allow always". Claude's Research mode can act without asking, so turn off those tools' editing options in the chat's tools menu before you start a Research run.

The report most teams rebuild by hand every close. Claude pulls the profit and loss and balance sheet from Xero or QuickBooks, lays the month against the last one with the basis printed at the top, and checks that last month's figures still match the pack you sent. What you get is a report you could send, plus the report settings behind every number. For the analysis that follows, see prompts for financial analysis.

Variance commentary is where AI saves finance the most time and where it does the most quiet damage, because an invented reason reads exactly like a real one. This workflow reads actuals by department from the ledger, reads the approved budget from your workbook, and keeps every explanation tied to an account, supplier or customer it can point to. QuickBooks keeps budgets on its Plus and Advanced plans (QuickBooks Help); most teams keep theirs in a workbook either way. FP&A teams building the forecast itself should see AI for FP&A.

Matching Stripe payouts to the bank is slow because a payout isn't a sale: it's a batch of charges, refunds, disputes and fees, dated by when Stripe expects it to arrive. Claude reads each payout's balance transactions from Stripe, matches the payouts to your bank statement export, and checks the fees against your ledger's Stripe clearing account. Anything it can't match stays on its own line.

A 13-week cash forecast is mostly bookkeeping: what's owed to you and when it will really arrive, what you owe and when it's due, plus payroll and tax. Claude builds it from Xero's or QuickBooks' open invoices and bills, times receipts from how each customer has paid before, and adds the fixed payments from your calendar. You get a weekly table and a chart, with every input listed.

Collections go faster when finance knows what sales knows: who owns the account, whether there's a renewal open, whether the customer has disputed the invoice. Claude reads the aged receivables from Xero or QuickBooks, looks up each overdue customer's owner and last contact in HubSpot, and drafts the reminders. Nothing gets sent: QuickBooks' connector can send reminders itself, so keep that tool switched off for this one.

Prompts reduce errors; they don't remove them. In a 2026 study of 14 models, the best produced working links more than 94% of the time, yet only 39 to 77% of their claims were accurate (Onweller et al., Cited but Not Verified, May 2026). In finance, the citation that opens is the ledger, and the claim that fails is usually the reason given for a number.

Common pitfalls

• Netting a difference away. A payout in transit, a missing fee and a duplicated receipt can cancel out. "Reconciled" needs every line matched, not a zero total.
• Reading a variance sign the wrong way. Spending £9,300 more than budget is a positive number and an unfavourable variance. Ask for the words, not just the sign.
• MRR in the revenue line. Billing tools normalise annual plans and leave out one-time charges. Recognised revenue comes from the ledger.
• Mixing bases. A cash-basis report for one month and accrual for the next makes a trend out of a setting.
• A tidy reason with nothing under it. "Delayed renewals" and "seasonal slowdown" are what commentary usually says. If the data doesn't show it, it's a guess.
• One entity read as the group. Xero's connector reads one organisation at a time, so a total from it is one company, not the consolidation.

Rules for the team

• No figure leaves finance until a person has traced it to the ledger. The report, not Claude's summary of it.
• Definitions live in the project. Change them there, once, not in individual chats.
• Read-only by default. Write access is switched on for a named job, then off again.
• Journals and provisions have a human owner. Claude drafts; the controller books.
• Financial data stays in the Team workspace, never a personal account.

One check before anything leaves finance

Count audits every answer automatically, inside the chat, and it checks more than the numbers. For each claim it asks: did this come from a source, and does the source really say it? Does the reasoning hold, or does a conclusion go further than the data? Was a definition, a basis or a sign convention assumed without saying so? Is a statement from the model's memory rather than anything in the chat? It replies with a canvas where each claim is labelled a citation, an assumption, model knowledge or a hallucination, with the reasoning chain under it. A variance read the wrong way, a payout called reconciled while still in transit and a cause no ledger shows all show up for what they are. Each workflow above shows an example.

How do I use AI for finance?

How much does it cost to set up AI for a finance team?

Is Claude good for finance?

Is Claude for Excel free?

Can Claude connect to NetSuite?

Can Claude connect to Xero and QuickBooks?

Is it safe to put financial data into Claude?

Will AI replace accountants?

Should finance teams use Claude or ChatGPT?